JPH1011583A - Class classification adaptation processor and learning device/method for class classification adaptation processing - Google Patents

Abstract

PROBLEM TO BE SOLVED: To learn a coefficient for improving the resolution of an input picture signal, an SN ratio and a luminance value and reducing compression distortion and to generate optimum data through the use of the coefficient. SOLUTION: In a processing circuit 3, the processing f(x) of LPF, compression expansion, a gain and offset is executed for the input picture signal. A class code (index) is generated from the picture signal to which the processing f(x) is executed and the generated class code is supplied to a learning circuit 7. Then, the picture signal to which the processing f(x) is executed, a teaching signal, the picture signal to which the processing f(x) is executed and the input picture signal, the input picture signal and the teacher signal are supplied to the learning circuit 7 by changing over switches 4 and 6. The learning circuit 7 executes learning by a least square method for obtaining the coefficient of a linear combination expression from the signals.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a classification adaptation method capable of further improving, for example, a resolution, a signal-to-noise ratio, a compression distortion and a luminance value of an image signal and / or an audio signal using a classification adaptation process. The present invention relates to a processing device, a learning device for class classification adaptive processing, and a learning method.

[0002]

2. Description of the Related Art Conventionally, as an application of class classification adaptive processing, SD (Standed Definition) to HD (Hi
gh Deginittion) image information conversion device, spatio-temporal model coding, MUSE image quality improvement, composite signal Y /
Various application ideas such as C separation have been proposed. That is, a pixel of a certain size of the spatiotemporal is divided into blocks, and this is subjected to some method, for example, ADRC (Ad
aptive Dynamic Range Coding)
Modeling is performed by linear linear combination for each class, that is, a prediction formula is established, and optimal data can be obtained by calculating the coefficients and pixels stored for each class. At this time, the coefficients stored for each class are obtained by learning using a least squares method or the like in advance.

[0003]

As described above, when processing such as resolution compensation is performed using the classification adaptive processing, a certain effect can be obtained. The present invention relates to an improvement of the above-described classification adaptive processing. That is, an object of the present invention is to perform some processing at the time of learning to obtain coefficients used in the classification adaptive processing, and to use the coefficients to further increase the resolution, SNR,
An object of the present invention is to provide a class classification adaptive processing device, a learning device for a class classification adaptive process, and a learning method capable of improving compression distortion and a luminance value.

[0004]

According to a first aspect of the present invention, a class generating means for generating class information from an input information signal in a class classification adaptive process for generating a predicted value or a correction value includes: A storage medium in which coefficients obtained by a learning process using the least squares method are stored for each of the
The coefficient is read from the storage medium in response to the class information,
Calculating means for generating a predicted value by an operation of a linear linear combination expression, wherein the coefficient performs processing on the input information signal, generates a class from the processed input information signal,
A class classification adaptive processing device characterized by being obtained by learning so as to minimize the square sum of an error between a processed input information signal and a teacher signal.

According to a third aspect of the present invention, there is provided a class classification in which a plurality of samples of an input information signal and a coefficient to be calculated are previously obtained for each class by learning in order to generate a prediction value or a correction value. In the learning device for adaptive processing, processing means for performing processing on the input information signal, class generating means for generating class information from the processed input information signal, and the processed input information signal, A learning device for class classification adaptive processing, comprising learning means for learning coefficients so that the sum of squares of the error with the teacher signal is minimized.

Further, according to a tenth aspect of the present invention, in order to generate a predicted value or a correction value, a plurality of samples of an input information signal and a coefficient to be calculated are previously obtained for each class by learning. In a learning method for adaptive processing, a step of performing processing on an input information signal; a step of generating class information from the processed input information signal; a processing of the input information signal; Learning a coefficient so that the sum of the squares of the error of the error is minimized.

According to the present invention, an LP is applied to an input image signal.
Processing f (x) such as F (low-pass filter) processing, compression / expansion, gain and / or offset is performed. In the case of processing f (x), for example, an LPF, the input image signal to which the LPF has been applied is reduced in sampling rate, for example, to ２, and a coefficient is obtained by a least square method with a teacher signal. In the classification adaptive processing using the obtained coefficients, the input signal is converted into a signal whose resolution is improved four times, and is output.

[0008]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing a general configuration of a learning device for class classification adaptive processing according to the present invention. 1
The input signal supplied from the input terminal indicated by
Input terminal B of switch 4 and input terminal D of switch 6
Supplied to The teacher signal supplied from the input terminal 2 is
It is supplied to the input terminal A of the switch 4. At this time, the input signal and the teacher signal may be the same signal as shown by the dotted lines in the figure, or a signal obtained by performing processing such as downsampling the teacher signal may be used.

In the processing circuit 3, the processing f (x) applied to the input signal includes, for example, LPF processing, compression / expansion, gain and / or offset, as described later. An example of the compression / decompression is an MPEG2 encoder and decoder. The correction target signal subjected to the processing f (x) is supplied to the class generation circuit 5 and the input terminal C of the switch 6. In the class generation circuit 5, the processing f
Class information (hereinafter, referred to as an index) is generated from the signal subjected to (x), and the index is supplied to the learning circuit 7.

In the above configuration, when the switch 4 selects the input terminal A and the switch 6 selects the input terminal C, the learning circuit 7 outputs the teacher signal from the input terminal 2 and
Learning by the least squares method is performed from the input signal that has been subjected to the processing f (x), and coefficients of a linear linear combination expression are obtained.
The coefficient is obtained by finding the minimum value of the square of the error between the teacher signal and the input signal subjected to the processing f (x). When the switch 4 selects the input terminal A and the switch 6 selects the input terminal D, the learning circuit 7
Learning is performed from the teacher signal from the input terminal 2 and the input signal from the input terminal 1. When the switch 4 selects the input terminal B and the switch 6 selects the input terminal C, the input is performed as the teacher signal. Input signal from terminal 1 and processing f (x)
The learning is performed from the input signal to which is applied.

The coefficients obtained by the learning circuit 7 are as follows:
The data is output from the output terminal 8 and stored in a storage medium (not shown) such as a memory. In FIG. 1, a class is generated from a signal obtained by performing some processing f (x) on an input signal,
In addition, it shows a general configuration for performing learning by the least square method from the input signal and the teacher signal. As described above, in this block diagram, when the input terminal A of the switch 4 and the input terminal D of the switch 6 are respectively selected, the path for learning using the teacher signal and the input signal is the same as in the related art. Secured.

FIG. 2 shows a general configuration of the classification adaptive processing using the obtained coefficients. Input terminal 1
The input signal supplied from 1 is supplied to the class generation circuit 12 and the delay (DL) circuit 14. Class generation circuit 1
In step 2, a class of the supplied input signal is generated, and the generated class is supplied to a coefficient memory 13 in which coefficients obtained by learning are stored. In the coefficient memory 13, the coefficient is read in response to the class from the class generation circuit 12, and the read coefficient is supplied to the prediction calculation circuit 15. In the prediction calculation circuit 15, the delay circuit 14 obtains optimal data by calculating the input signal delayed by a predetermined time and a coefficient.

In the above-described general configuration of the present invention, the processing circuit 3 performs the LPF processing, thereby increasing the resolution according to a first embodiment shown in FIG.
This will be described with reference to FIG. First, video signals captured by the imaging system having two different resolutions for the same subject are used. At this time, the input signal obtained by the imaging system having the lower resolution is converted from an analog signal to a digital signal by an A / D converter having a sampling frequency of 13.5 MHz, for example, and this video signal is represented as LS. The teacher signal obtained by the higher resolution imaging system is, for example, 27.
With an A / D converter with a sampling frequency of 0 MHz,
The analog signal is converted to a digital signal, and this video signal is represented as HS.

As described above, using the input signal (video signal LS) and the teacher signal (video signal HS), for example, in FIG. 1, the input terminal A of the switch 4 and the input terminal D of the switch 6 are connected. By learning in a state in which the signal is sampled, a coefficient for generating a signal having a twice as high sampling rate from a signal having a low sampling rate is calculated.

By learning while the input terminal A of the switch 4 and the input terminal C of the switch 6 in FIG. 1 are connected, a signal having a four-fold higher sampling rate is generated from a signal having a lower sampling rate. Is calculated. This is because, for example, LPF processing is performed as a process f (x) on a coefficient generated by learning on the input signal, and the number of pixels of the supplied input signal is not reduced, and only the band of the input signal is reduced. This is because, as shown in FIG. 3A, a signal (LPF-LS) whose sampling rate is reduced to half, for example, and a teacher signal are used.

More specifically, the input signal LS (13.5
(Mhz) signal LPF-LS
(6.75 MHz) and the learning signal HS (27.0 MHz), using the obtained coefficients, as shown in FIG. 3B, the input signal LS (1
It is possible to generate a signal (54.0 MHz) having a sampling rate four times higher than that of 3.5 MHz.

In the conventional learning method, it was only possible to obtain twice the resolution of the input signal from the input signal. However, the resolution was obtained by using the coefficient obtained by the learning method to which this embodiment is applied. When performing the raising process, it is possible to create a signal having a resolution four times that of an input signal having a normal sampling rate.

Next, as a second embodiment, learning when the SN ratio is improved will be described. Generally, when a video signal captured by a camera passes through a transmission system or a recording / reproducing system, some noise is added to the video signal, and the S / N ratio deteriorates. Normally, a coefficient for improving the S / N ratio is learned by using a camera output as a teacher signal and a signal passed through a transmission system or a recording / reproducing system as an input signal. When performing SN improvement using this coefficient, when an unexpected amount of noise is added to the input signal,
We cannot hope for a significant improvement.

Therefore, in the second embodiment, a random number capable of controlling characteristics such as the magnitude and the frequency of occurrence is generated by a computer, for example, as a process f (x), and added to the input signal output from the camera. Thus, a signal contaminated with artificial noise is created, and a coefficient for improving the SN ratio is calculated by learning with a teacher signal. By doing so, it is possible to easily control the learning of the coefficient for the assumed noise characteristic, and it is possible to improve the improvement effect. It is expected that the effect of improving the S / N ratio will be significantly improved by learning together with the signal transmitted through the conventional transmission system and recording / reproducing system.

Next, as a third embodiment, learning in the case of improving compression distortion will be described. In the case of transmitting an image via the Internet or the like, the image is usually transmitted after being compressed, so that the image reproduced on a computer monitor often has a unique distortion. In TV broadcasting, not only uncompressed image signals but also MPE
Images compressed by G or the like are also being handled, and image distortion due to compression has become a problem.

The image distortion due to the compression generally differs depending on the compression method. Thus, in the third embodiment, the input image signal is subjected to compression / expansion processing of various assumed compression methods as processing f (x), and
Classification adaptive processing using acquired coefficients for any unknown input signal by learning coefficients for improving compression distortion using uncompressed image signals that are teacher signals Is performed, it is possible to improve compression distortion.

Next, as a fourth embodiment, learning when performing luminance correction will be described. FIG. 4 shows a more specific block diagram of the learning device for class classification adaptive processing shown in FIG. 1 for luminance correction. The luminance signal Y of the image signals supplied from the input terminal 21 is converted from an analog signal to a digital signal by the A / D converter 22. In the A / D converter 22, for example, when sampling is performed with a clock of 13.5 MHz, the size of an image is about 720 pixels horizontally × 480 lines vertically in one frame. The luminance signal converted to a digital signal is
The signal is supplied from the A / D converter 22 to the learning circuit 23 and the gain / offset circuit 24.

In the gain / offset circuit 24, processing for changing the gain and / or offset of the supplied luminance signal is executed as processing f (x). The gain is related to the brightness (contrast) of the lighting, and the offset is related to the average brightness of the lighting. At this time, the gain need not always be 1;
A value such as 9, 1.2 may be used. The luminance signal subjected to the gain and / or offset processing is supplied to the average / standard deviation circuit 25 and the delay circuit 27.

In the average value / standard deviation circuit 25, for example, one field or one of the supplied luminance values is supplied as described later.
The average and standard deviation per frame are determined. The average value / standard deviation circuit 25 has a table for obtaining a frequency distribution for each luminance value, and an average value of luminance is calculated from the frequency distribution multiplied by one field period or one frame period as shown in FIG. In addition, the standard deviation is calculated. The calculated average value and standard deviation are supplied from the average value / standard deviation circuit 25 to the quantization (Q) circuit 26. Formula (1) shows a calculation formula for calculating the average value of luminance, and Formula (2) shows a calculation formula for calculating the standard deviation.

Average value = Σ (brightness value × frequency) / total frequency (1) Standard deviation = √ (Σ (luminance value−average value) ² × frequency) / total frequency) (2) where √ () is The calculation result in parentheses is the square root.

In the quantization circuit 26, the calculated average value and standard deviation are quantized by a bits and b bits, respectively, to generate a code of total n bits (n = a + b). The n-bit code is supplied from the quantization circuit 26 to the degenerate ROM 32. Further, the n-bit code is a pattern of a so-called luminance distribution, and by looking at this, it is determined whether the luminance distribution is dark or bright, and whether the luminance distribution is flat or steep. can do.

On the other hand, in the delay circuit 27, a delay is performed for a time (one field or one frame + α) until an n-bit code is generated, and the output is sent to the blocking circuits 28 and 30 and the delay circuit 33. Supplied to In the blocking circuit 28, a plurality of pixels in the space around the target pixel are selected and blocked. The blocked pixels are supplied to the ADRC circuit 29.
In the ADRC circuit 29, the maximum value and the minimum value are selected from the blocked pixels as described later, and each pixel is requantized to generate an m-bit code.
2. This code is a pattern in which the so-called spatial luminance change is performed. The class classification based on the m-bit code is based on S / S rather than the brightness correction itself.
This has the effect of improving the N ratio, improving the resolution, and the like.

In the blocking circuit 30, a plurality of pixels in the space around the pixel of interest to be corrected are selected and blocked. The blocking performed by the blocking circuit 30 and the blocking circuit 2 described above
8 may differ from the blocking performed. That is, there is no problem even if the pixel selected in the blocking circuit 28 is different from the pixel selected in the blocking circuit 30. The blocked pixels are supplied from the blocking circuit 30 to the averaging circuit 31. In the averaging circuit 31, the average value of the luminance near the pixel of interest is calculated, and the calculated average value is shifted, and p (<
8) The data is quantized into bits and supplied to the degenerate ROM 32.

As described above, the average value of the luminance of each of the blocked pixels is calculated, that is, by setting the luminance level as one of the class classifications, the manner of correction in the level direction can be varied. . For example, it is possible to correct only a bright part or a dark part, or to perform correction in consideration of the γ characteristic. In addition, the function of averaging the luminance levels also serves to prevent the luminance correction from being overly sensitive.

In the above description, three types of class classification codes are generated. If these are simply combined, the number of classifications becomes enormous, and the capacity of a coefficient ROM described later becomes enormous. Therefore, the n-bit ADRC circuit 2 from the quantization circuit 26
The m bits from 9 and the p bits from the averaging circuit 31 degenerate the supplied bits in the degeneration ROM 32. Specifically, in the degenerate ROM 32, a class code (index) reduced from (n + m + p) bits to q bits is generated in order to degenerate the class. In this way, a q-bit class code is finally generated from the degenerate ROM 32, and the q-bit class code is supplied to the learning circuit 23.

Although the details of the method of degeneracy will not be described here, from the coefficient sets corresponding to all the classes learned with (n + m + p) bits, the ones with small inter-coefficient norms are summarized as vector quantization techniques. A method of degenerating shall be used. That is, a distance (norm between coefficients) of a corresponding coefficient is obtained between two coefficient sets, and the coefficient sets are put together based on the distance.

The output of the delay circuit 33 for which the delay adjustment is performed is supplied to a blocking circuit 34, and the blocking circuit 34 blocks a plurality of pixels around the target pixel. Each of the blocked pixel values is supplied to the learning circuit 23.

Then, as described above, the A / D converter 2
2 are supplied to the learning circuit 23. The learning circuit 23 forms an n-tap linear temporary combination model, and the respective coefficients are calculated by the learning circuit 23. Each calculated coefficient is taken out from the output terminal 35 and the coefficient R
Stored in OM. The learning circuit 23 learns coefficients for each class by the least square method described later. The coefficient obtained by the learning is output via the output terminal 35 and stored in a storage medium such as a memory for each class.

Here, an example of the average value / standard deviation circuit 25 will be described with reference to FIG. A luminance value is supplied from an input terminal 41. The supplied luminance value is stored in the luminance frequency distribution table 4.
2 and generates a frequency distribution table of the luminance levels in one field or one frame in the luminance frequency distribution table 42, for example. Based on the generated table, the average value is calculated by the average value calculation circuit 43 according to the equation (1).
Is calculated, and the calculated average value is supplied to the standard deviation calculation circuit 44 and taken out from the output terminal 45. In the standard deviation calculating circuit 44, the standard deviation is calculated from the frequency distribution table and the average value by the equation (2), and the calculated standard deviation is taken out from the output terminal 46. When the extracted standard deviation is small, the width of the frequency distribution is narrow, and when the standard deviation is large, the width of the frequency distribution is wide.

Here, an example of the configuration of the ADRC circuit 29 will be described with reference to FIG. Blocked data is supplied from the input terminal 51. The supplied data is supplied to a maximum value detection circuit 52, a minimum value detection circuit 53, and a delay circuit 54. The maximum value detection circuit 52 detects the maximum value of the pixel value in the block, and the minimum value detection circuit 5
At 3, the smallest value of the pixel values in the block is detected. The subtractor 55 subtracts the minimum value from the maximum value and calculates the dynamic range DR of the block. The calculated dynamic range DR is supplied to the adaptive requantization circuit 57.

In the delay circuit 54, the maximum value detection circuit 52 and the minimum value detection circuit 53 are each subjected to a time delay for detection, and are output one pixel at a time. In the subtractor 56,
The minimum value is subtracted from each of the blocked pixels, and the subtracted value is supplied to the adaptive requantization circuit 57. The adaptive requantization circuit 57 quantizes the subtraction value for each pixel using a predetermined quantization step width according to the dynamic range DR. In the parallelization circuit 58, the quantized pixels are parallelized in block units and output as coded data from an output terminal 59.

As a learning method, a least-squares method is used to determine the relationship between the pixel values of a large number of input signals for correction and the pixel values of a teacher image signal. First, it is assumed that there is a linear linear combination between the above-described values, and a linear linear combination model is shown below.

Linear linear combination model: (observation equation) XW = Y (3)

(Equation 1)

Release by the method of least squares: (residual equation)

(Equation 2)

From equation (5), to find the most probable value of each w _i ,

(Equation 3) The condition that minimizes

[0041]

(Equation 4) , W ₁ , w ₂ ,
... Find out w _N. From equation (5),

[0042]

(Equation 5) When the condition of Expression (6) is established for i = 1, 2,..., N,

[0043]

(Equation 6) Is obtained. Here, the following normal equation is obtained from the equations (5) and (8).

[0044]

(Equation 7) Since this is a simultaneous equation having exactly N number of unknowns, each w _i that is the most probable value can be obtained.

To be more precise, in equation (9), w _i

(Equation 8) Can be solved if the matrix is regular. (Where k = 1, 2,..., N, l = 1, 2,.
N) Actually, simultaneous equations are solved using the Gauss-Jordan elimination method (sweep-out method).

Next, FIG. 8 shows a block diagram of hardware for performing the operation of the least squares method. In the learning block diagram of FIG. 4, the pixel values x _{1 to} x _{N of} the block centering on the pixel to be corrected and the teacher pixel value δ corresponding to the pixel
y is input and the class code (index)
Is entered. The circuit of the least square method is roughly divided into a normal equation generation circuit 61 and a CPU 62. The normal equation generation circuit 61 includes a multiplier array 63, an addition memory 6
4 and a decoding unit 65. The CPU 62 includes, for example, a CPU that performs an operation of a sweeping-out method to obtain a coefficient. The multiplier array 63 has one set of memories (or registers) for the target pixel position, and the addition memory 64 has as many sets of memories (or registers) as the number of classes. In the decoding unit 65, the supplied class code (index) is decoded.

Here, the multiplier array 63 will be described with reference to FIG.
This will be described with reference to FIG. The teacher pixel value δy corresponding to the pixel value of the block around the pixel to be corrected is multiplied by each element in the multiplier array 63 of the normal equation generation circuit 61 as shown in the figure. The result is supplied to the addition memory 64.

The addition memory 64 comprises an adder array 71 and memory (or register) arrays 72 _{1 to} 72 _N as shown in FIG. The adder array 71, the output from the result memory (or register) array 72 ₁ to 72 _N of the multiplier array 63 are supplied. The addition result is output from the adder array 71 memory (or registers) in the array 72 ₁ to 72 _N. At this time, which memory (or register) array 7
Or 2 ₁ to 72 _N is selected, the class code supplied to the decoding unit 65 (index) is uniquely determined by being decoded. That is, the memory (or register) array 72 is selected for each class determined by the index. The result of the product-sum operation is updated in the selected memory (or register) array 72,
It is memorized.

Note that the position of each array is related to w _i of the normal equation (9).

(Equation 9) Corresponding to the position. As can be seen from the normal equation (9), if the upper right term is inverted, it becomes the same as the lower left, so that each array has a triangular shape.

As described above, a product-sum operation is performed for each class during a certain period, and a normal equation for each class is generated for each pixel position. The result of each term of the normal equation for each class is stored in the memory (or register) array corresponding to each class, and then each term of the normal equation for each class is supplied to the calculation circuit of the sweeping method. Is done. This calculation is performed by the CPU 62.
The calculated set of coefficients is used by being written into a coefficient table constituted by a coefficient ROM.

Next, a luminance correction circuit capable of improving the resolution of an image signal by using the coefficients obtained by the above-described learning will be described with reference to FIG. In describing this example, the same portions as those in the above-described fourth embodiment are denoted by the same reference numerals, and description thereof will be omitted.

The luminance signal Y supplied from the input terminal 81
Is supplied to the A / D converter 82, where the A / D converter 82 samples at, for example, 13.5 MHz and digitized signals are output to the average value / standard deviation circuit 25 and the delay circuit 27. . In the degenerate ROM 32, the n-bit code from the quantization circuit 26 and the ADRC circuit 29
, And the p-bit code from the averaging circuit 31 are degenerated, and the q-bit class code (index) is supplied to the coefficient ROM 83. In the coefficient ROMs 83 _{1 to} 83 _N , addressing is performed using the supplied class code, and the coefficients w _{1 to} w _N are read. The read coefficients w _{1 to} w _N are respectively
It is supplied to the 4 ₁ -84 _N.

In the blocking circuit 34, a plurality of pixels around the target pixel are blocked. Each of the blocked pixel values is supplied to multipliers 84 _{1 to} 84 _N. The multiplier 84 ₁ -84 _N, and the coefficients w ₁ to w _N from the coefficient ROM 83 ₁ to 83 _N, each pixel value is blocked are multiplied and the multiplied value is supplied to the adder 85. Adder 8
At 5, the multiplied values from the multipliers 84 _{1 to} 84 _N are added. That is, the multipliers 84 _{1 to} 84 _N and the adder 85
, A luminance correction value is predicted by performing a product-sum operation. The predicted value is calculated by the D / A converter 86 as D
/ A conversion and output terminal 87 as corrected luminance value Y '.
Taken out of

[0054] coefficient ROM 83 ₁ to 83 coefficients w ₁ to w _N to be read from the _N, the sampling rate is 6.75MHz when learning the aforementioned signal LPF-LS and 27.0FM
This is a coefficient obtained based on the Hz video signal. By using this factor when generating predictions,
The luminance signal Y having a sampling rate of 13.5 MHz is 4
It is possible to generate a luminance signal Y ′ having a sampling rate of 54.0 MHz, which is twice as high.

In this embodiment, the method of realizing all the hardware is described. However, the calculation may be performed by software by taking digitized data into a computer.

In this embodiment, the resolution, S / N
Processing f (x) to be applied to an input signal when learning is performed to obtain a coefficient for improving the ratio, compression distortion, and luminance value
Are performed separately, but two or more of these processes may be combined. For example, by performing LPF processing on an input signal after gain and / or offset,
The obtained coefficients can be used to further improve the resolution and brightness values.

[0057]

According to the present invention, the input signal is L
Learning is performed using a signal generated by performing processing such as PF, compression / expansion, gain and / or offset, and a teacher signal, thereby obtaining coefficients. By the classification adaptive processing using the coefficients, it is possible to further improve the resolution, the SN ratio, the compression distortion, and the luminance value.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a general configuration example of a learning device for class classification adaptive processing according to the present invention.

FIG. 2 is a block diagram showing a general configuration example of an embodiment of a class classification adaptive process according to the present invention;

FIG. 3 is a schematic diagram of an example used for explaining the resolution of the present invention.

FIG. 4 is an embodiment of a learning circuit corresponding to a luminance correction device to which the present invention can be applied.

FIG. 5 is a frequency distribution table used for describing an embodiment to which the present invention can be applied.

FIG. 6 is an example of an average value / standard deviation circuit used in a luminance correction device to which the present invention is applied.

FIG. 7 is an example of an ADRC circuit used in a luminance correction device to which the present invention is applied.

FIG. 8 is a block diagram illustrating an example of a least squares method applied to the learning circuit according to the present invention.

FIG. 9 is a schematic diagram illustrating an example of a multiplier array according to the present invention;

FIG. 10 is a schematic diagram illustrating an example of an addition memory according to the present invention;

FIG. 11 is a block diagram showing an embodiment of a luminance correction device to which the present invention can be applied.

[Explanation of symbols]

3 Processing circuit 4, 6 Switch, 5 Class generation circuit, 7 Learning circuit

Claims (11)

[Claims] In a class classification adaptive process for generating a predicted value or a correction value, a class generating means for generating class information from an input information signal, and a coefficient previously obtained by a learning process by a least square method for each class. And a calculating means for reading out the coefficient from the storage medium in response to the class information and generating a predicted value by a calculation of a linear linear combination expression, wherein the coefficient is Processing the input information signal, generating a class from the processed input information signal, and minimizing the sum of squares of the error between the processed input information signal and the teacher signal. A classification adaptive processing device obtained by learning. 2. The adaptive classification processing apparatus according to claim 1, wherein the input information signal is generated from the teacher signal. 3. A learning apparatus for class classification adaptive processing in which a plurality of samples of an input information signal and a coefficient to be calculated are previously obtained for each class by learning in order to generate a prediction value or a correction value. Processing means for performing processing on the information signal; class generating means for generating class information from the input information signal subjected to the processing; and square of an error between the input information signal subjected to the processing and the teacher signal A learning device for class classification adaptive processing, comprising: learning means for learning coefficients so as to minimize the sum. 4. The learning device for class classification adaptive processing according to claim 3, wherein the input information signal is generated from the teacher signal. 5. The learning apparatus for class classification adaptive processing according to claim 3, wherein first selection means for switching between the teacher signal and the input information signal, and the input information signal and the processing are performed. A second selecting means for switching between an input information signal and an output of the first selecting means, and an output of the second selecting means being supplied to the learning means. Learning device for classification adaptive processing. 6. The learning apparatus for class classification adaptive processing according to claim 3, wherein the processing is low-pass filtering processing on the input information signal, and learns a coefficient for increasing the resolution. A learning device for class classification adaptive processing, characterized in that: 7. The learning device for class classification adaptive processing according to claim 3, wherein the processing is to add a t-random noise to the input information signal by generating a random number, and to improve an SN ratio. A learning device for class classification adaptive processing, characterized in that the learning device learns the coefficient of. 8. The learning device for class classification adaptive processing according to claim 3, wherein the processing is to add distortion due to compression / expansion processing to the input information signal, and to improve distortion due to compression. A learning device for class classification adaptive processing, characterized in that the learning device learns the coefficient of. 9. The learning device for class classification adaptive processing according to claim 3, wherein the processing is a gain and / or offset processing on an input information signal, and learns a coefficient for improving a luminance distribution. A learning device for class classification adaptive processing, characterized in that: 10. A learning method for class classification adaptive processing in which a plurality of samples of an input information signal and a coefficient to be calculated are previously obtained for each class by learning in order to generate a predicted value or a correction value. Processing the information signal; generating class information from the processed input information signal; and minimizing the sum of squares of the error between the processed input information signal and the teacher signal. Learning a coefficient so as to obtain a classifying adaptive processing. 11. The learning method for class classification adaptive processing according to claim 10, wherein the input information signal is generated from the teacher signal.